Personalised feedback that actually lands

The most common AI-feedback failure

Ask any chatbot, in any classroom-feedback context, “give feedback on this student’s essay,” and you will get a recognisable shape: an opening sentence of warm praise, two or three paragraphs of general affirmation (“you have shown good understanding of the topic”, “your effort is appreciated”, “the structure is clear and easy to follow”), one paragraph of vague suggestion (“you could expand on some points”, “consider adding more examples”), and a closing line of encouragement. Two hundred words. Fluent. Polite. Useless.

It is useless for a specific reason. It names nothing. Nothing the student actually wrote is quoted. Nothing the student specifically did well is pinned to a sentence in their work. Nothing the student should change next time is concrete enough to act on. The feedback could be pasted, almost word for word, onto any other essay in the class, and would be roughly as accurate — and roughly as inert.

A child reads this paragraph, registers that the teacher (or the teacher’s machine) is broadly pleased, and learns nothing about what specifically to do on the next piece of work. The professional cost is hidden because everyone smiles. The pedagogical cost is total.

The structured feedback rubric

The fix is not “ask the model to be more specific.” The model, asked to be more specific, will produce a longer paragraph of slightly more elaborate generalities. The fix is to give the model a structure that has no room for generic praise.

A serviceable structure — three lines, full stop — looks like this:

One specific strength, with a direct quote from the student’s own work and one sentence explaining why that move was effective. “When you wrote the earthquake of 1934 exposed the limits of the Rana administration's reach, you connected a single event to a structural argument — that is the move that turns a description into history.”

One specific improvement, with a clear next step the student can act on the next time they write. “Your second paragraph names three causes but does not say which one mattered most or why; next time, end the paragraph with one sentence ranking them.”

One self-question the student can ask themselves the next time they sit down to write. “Before you start your next history paragraph, ask: ‘which of my causes was the most decisive, and what evidence do I have for that?’”

No general praise. No “well done overall.” No closing encouragement. Three lines, each one doing distinct work, all three anchored in the student’s specific paper. A child who reads this knows what they did, what to change, and what to think about next time.

The prompt that produces it

Here is the prompt that gets the structure reliably. Paste this, replace the placeholders, and read what comes back.

“You are providing feedback to a Class 9 student in Nepal on a short-answer history response. Do not write general praise. Do not write a closing paragraph of encouragement. Use exactly this three-line structure: (1) ONE specific strength, with a direct quote from the student’s work and one sentence explaining why it was effective; (2) ONE specific improvement, with a clear, concrete next step the student can take on the next piece of work; (3) ONE self-question the student can ask themselves before they next write a paragraph of this kind. Total length: under 100 words. Tone: respectful, direct, peer-to-peer, not patronising. Student’s response is pasted below.”

The first time you run this, the model may still try to slip in a “great work overall” sentence at the end. Tell it again, more sharply: “Remove the closing encouragement. The three lines stand alone.” By the third or fourth prompt of the day, it stays in the structure.

Bilingual feedback, because the child reads in one language

The unfair truth about feedback in Nepali classrooms is that the children who most need detailed comments are often the children whose English is weakest, and detailed feedback in English bounces off them the way it always has. The richer the feedback, the bigger the comprehension gap. AI changes the arithmetic here in a way that almost nothing else has.

Two rules of thumb. When the student is more comfortable reading Nepali, the feedback is in Nepali, even if the assignment was written in English. Better that a child reads three precise lines than skims six polite ones. When the work itself is an English-language exercise — composition, comprehension, a +2 English paper — the feedback is in English, because the medium is part of the learning. For the middle case, where the student is building reading capacity in English, the model can produce both: the structured three lines in English, with a one-line Nepali summary underneath that names the single thing to change.

The structured prompt above takes one extra clause: “Provide the three lines in Nepali in respectful student-register, with no English calques for terms like ‘thesis’ or ‘paragraph’ — use the natural Nepali equivalents.” Read the Nepali version before sending. Honorifics and student register are the places the model is weakest in Nepali, and a slightly off tone on a feedback note lands worse than the same tone in a report.

The reading-aloud test

You will need a quick check to know whether a given piece of AI feedback is doing its job or sliding back into the generic-praise shape. The test is small and reliable: read the feedback aloud in your own voice, slowly, as if you were saying it to the child in front of you after class. Would the student stop and re-read it? Would they nod and remember a specific thing they wrote? Would they know what to do next week?

If yes, the feedback is specific enough. Send it. If no — if the words feel polite and smooth and could be about any essay — go back to the prompt. The line that failed is almost always the “specific strength” line, because that is where the model is most tempted to fall back on warmth. Insist on the quote.

A short word on workload

The honest reason teachers in Nepal write generic feedback is not that they do not care. It is that forty-five papers times five minutes of thoughtful comment per paper is a weekend, every week, on top of teaching. The reason this section matters is not that AI writes better feedback than a teacher who has all the time in the world; it does not. The reason this matters is that AI writes good-enough feedback faster than a teacher who has no time at all — and good-enough specific feedback is dramatically better than thorough generic feedback. The trade is honest. Take it.

Check your understanding

Quick check

—

True or false: a long, warmly worded AI-generated paragraph of general praise is good feedback for a struggling student because it builds confidence.

1. ATrue — encouragement is the most important part of feedback.
2. BFalse — generic praise names nothing the student actually did, gives them nothing specific to change, and is typically skimmed and discarded; the fix is a three-line structure with a quoted strength, a concrete next step, and a self-question.
3. CTrue — as long as the praise is in the student’s preferred language.
4. DFalse — but only because the paragraph is too long; the same praise in two sentences would work.

Check answer→

What comes next

Aligned items, defensible rubric grading, and feedback that lands close the loop on assessment. The next chapter steps back from grading and into the medium itself: language, translation, and the bilingual classroom. The children in front of you are reading, thinking, and arguing in Nepali, in English, and increasingly in a mixed register that neither textbook fully covers — and AI’s relationship with Nepali is, as you will see, both more useful and more fragile than the marketing suggests.